Sticker Shock: AI Demand Nearly Quadruples DDR5 Memory Prices
A RAM kit that cost $100 a year ago now runs $375 as AI data centers eat the memory supply chain, while a showroom greeter robot and a 15,000-unit production run illustrate two very different meanings of humanoid robot "deployment."
Capability & Integration
- Z.ai released GLM-5.3 on August 14, reusing the same 743B-parameter base weights as GLM-5.2, with every reported gain coming from post-training rather than retraining. The company reports a jump from 4.6% to 28.3% on Terminal-Bench 3.0 and an 84.5% score on its CyberGym security benchmark — company-reported, not independently verified — per MarkTechPost. Weights aren’t public yet; Z.ai says it will release them roughly two weeks after launch.
Review
“84.5% score on its CyberGym security benchmark” — single-sourced, company-self-reported numeric claim, structurally identical to the AgiBot claim below that gets a formal unverified callout, but here it only gets an inline parenthetical — suggested action: add unverified flag for consistency.
- Anthropic opened a public beta letting Claude Code sessions run on infrastructure customers control rather than Anthropic’s own servers, available to Team and Enterprise plans since August 6. Repository checkouts, build artifacts, and secrets stay on the customer’s machines, though model inference itself still runs through Anthropic, per Claude.
Robotics
- BYD unveiled its first humanoid robot, “Xiao Di,” at a Zhengzhou showroom in early August — a 1.61-meter service robot meant to greet customers and translate between languages, with plans for two or three units per store. This is a public interaction demo aimed at retail theater, not a production deployment, per South China Morning Post.
- AgiBot’s cumulative humanoid production reportedly passed 15,000 units in August, per Gasgoo.
Unverified
Figure comes from Chinese trade press citing the company’s own count; how many units are doing paid work versus sitting in inventory isn’t disclosed — the same caveat that applied to Figure AI’s 1,000-unit milestone last week.
Review
“passed 15,000 units in August” — this milestone happened June 28, 2026 (per Gasgoo’s own article metadata, AgiBot’s own press release, and independent trade coverage), not August; placing it under today’s news misdates a ~7-week-old event — suggested action: fix (correct to “in June”) or cut.
Hardware & Supply Chain
- TSMC’s board approved roughly $29.4 billion in capital spending for advanced nodes, packaging, and fab construction, alongside a confirmed joint venture with Sony to build a ¥747 billion image-sensor plant in Kumamoto, targeting volume production in 2029, per Sony Semiconductor Solutions.
- A mainstream 32GB DDR5 kit that cost about $100 a year ago now sells for a minimum of $375, as AI data centers pull memory-fab capacity toward HBM — which consumes roughly three times the wafer capacity per bit of standard DDR5 — and away from consumer parts, per Tom’s Hardware.
Environmental & Cultural Impact
- Update: Amazon disclosed for the first time that its data centers used 2.5 billion gallons of water worldwide last year — about 5% of what metro Seattle consumes annually — a week after Google reported 10.9 billion gallons, per Yahoo Finance. Investor pressure has since ranked the hyperscalers on transparency: Meta first, Amazon last, per Tom’s Hardware.
Review
“a week after Google reported 10.9 billion gallons” — sequencing appears backwards: Amazon’s disclosure was published June 11, 2026, roughly three weeks before Google’s June 30 report, not a week after it — suggested action: fix (correct the sequence/timing).
Review
“Investor pressure has since ranked the hyperscalers on transparency: Meta first, Amazon last” — the cited Tom’s Hardware article is from April 2026, predates both disclosures it’s framed as following, and its content describes inconsistent disclosure formats without an explicit ordinal ranking — likely claim-link mismatch or fabricated detail — suggested action: verify against full article text; likely cut or rewrite without the ranking claim.
- The EU began enforcing AI Act transparency rules on August 2, requiring deepfakes to carry disclosure labels and chatbots to identify themselves as automated at first contact; violations carry fines up to €15 million or 3% of global annual turnover, per the European Commission.
AI in the Wild
The Motion Picture Association and ByteDance reached an agreement on AI guardrails covering Seedance, Seedream, TikTok, CapCut, and Dreamina, six months after an AI-generated fight video depicting Tom Cruise and Brad Pitt went viral and prompted the MPA’s first-ever cease-and-desist to a generative AI company. The clip’s origin was never disputed — it was made with ByteDance’s Seedance 2.0 — per Deadline. Its resolution shows the fastest-moving fight over AI content isn’t disinformation; it’s unauthorized likeness and IP, settled by negotiated memorandum rather than by law.
Takeaway
Takeaway
Today’s stories split cleanly along what AI has made cheap versus what it hasn’t. Shipping a coding model, standing up a self-hosted agent, generating a convincing celebrity deepfake — all of that moves at software speed. What doesn’t: a RAM stick costs nearly four times what it did a year ago, an image-sensor plant needs until 2029 to reach volume production, and a robot that greets showroom visitors is a fundamentally different machine from one logging shift hours on a line. The physical layer underneath AI isn’t catching up to the software layer — the gap is the story now, not a footnote to it.